Homeland Security Challenge
During disasters, people are increasingly turning to social networks to communicate about challenges they face in affected areas. These networks could provide a goldmine of information about the effectiveness of current disaster response efforts and published guidelines; however, these complex, large-scale, socio-technical systems do not readily lend themselves to analysis or evaluation. This can make it difficult to identify the ad hoc human networks built and utilized during the recovery process, further complicating attempts to compare actual relief efforts against national guidelines, such as the National Response Framework (NRF) outlined by the U.S. Department of Homeland Security. Without this data, recommendations for improvement of disaster response policy may be uninformed, losing opportunities to increase the effectiveness and resilience of these efforts.
The CIRI team is leveraging methods and theories from natural language processing (NLP) and social network analysis (SNA) to collect and analyze user generated and professionally generated text data to outline the multi-modal networks that were formed within prior emergency response efforts. From this, they will develop a classification schema and method of labeling network links for ease of inference. This will enable the team to compare findings with current national disaster response guidelines to determine if efforts and results on the ground align with published guidelines, discover what actions humanitarian assistance and disaster response (HADR) agencies are willing and able to perform above and beyond national guidelines, and shed light on what national guideline activities are absent in the actual response data.
- Department of Homeland Security and Critical Infrastructure Resilience Institute, 2019-2020
The Humanitarian Assistance and Disaster Relief (HADR) project has endeavored to use natural language processing (NLP) to increase situational awareness of humanitarian relief operations. We based our efforts on prior work in and across a number of fields heavily engaged in similar projects. These fields include crisis informatics, humanitarian response, and information systems for crisis response and management respectively. Several prior studies have proposed the use of text mining (TM) and NLP algorithms to efficiently summarize information and aid HADR efforts. Our study aims to assess the difference in situational awareness information based on choices made in developing these algorithms. In particular, we assess the impact of the choice of data source, method, and method implementation. We also compare the outputs of these algorithms against an established ground truth comprised of official situational reports. We select three TM/NLP methods highly relevant to HADR efforts: topic modeling, document summarization, and needs detection. Our methods-oriented approach augments existing efforts by comparing different sources of text-based communication to an established ground truth or what is already known to be happening on the ground. By comparing disparate text-based sources to what we already know, new items can be added as they appear. Our current efforts surround what sources of information represent the most reliably useful information to expand situational awareness for everyone involved in the disrupted region.
- Sarol, J., Dinh, L., and Diesner, J. (2021). Information biases in situational awareness of crisis events due to data, method, and implementation choices for text mining. Proceedings of International AAAI Conference on Web and Social Media (ICWSM) 2021. Venice, Italy.
- Dinh, L., Sarol, J., & Diesner, J. (2020). How does situational awareness of emergencies depend on choices about data sources, analysis methods, and implementation of algorithms? Poster at 6th Annual International Conference on Computational Social Science (IC2S2), Cambridge, MA.
- Department of Homeland Security and Critical Infrastructure Resilience Institute, 2017-2018